Talk by Bill Sprangue of UC Berkeley. Given to the Redwood Center for Theoretical Neuroscience at UC Berkeley.
Abstract The efficient coding hypothesis broadly predicts that the tuning properties of neurons should reflect the statistics of the signal being encoded. For example, a sparse coding representation of natural images gives rise to receptive fields similar to those found in V1 (Olshausen and Field, 1996). A common corollary of this idea is that higher-level psychophysical phenomena can be explained by regularities in the visual environment. We asked if the statistics of disparity during natural viewing could provide a basis for understanding some phenomena in stereovision and the distribution of disparity preferences in binocular V1 neurons. To answer this question, we built a mobile binocular eyetracker that simultaneously measures binocular fixations and the 3D structure of the scene while subjects perform everyday activities such as making a sandwich, socializing, and navigating an environment. From these data we reconstructed the retinal images and disparities experienced by subjects as they performed these natural tasks. We found that disparities encountered in natural viewing have clear regularities that vary across tasks. Horizontal disparities tend to be uncrossed above fixation and crossed below, similar to the horopter - i.e. the region of space in which stereopsis is the most precise. Vertical disparities tend to be positive in the upper-left and lower-right visual field and negative in the other quadrants. We also compared this pattern to the distribution of preferred disparities in 820 V1 neurons from five single-unit studies: Cumming (2002), Durand, Celebrini, and Trotter (2007), Gonzalez, Bermudez, Vicente, and Romero (2010), Prince, Pointon, Cumming, and Parker (2002), and Samond, Potetz, and Lee (2012). We found that preferred disparities reflect the statistics of disparity in different regions of the visual field. Neurons from the upper visual field preferred uncrossed disparity, while neurons from the lower visual field preferred crossed disparities. Our analysis of V1 neurons is consistent with the hypothesis that the distribution of preferred disparities reflects the statistics of disparity across the visual field and that the region of best stereovision is well adapted to the pattern of disparity experienced in natural viewing.